The competition isn't even close to SORA.

@martin231181

3 ай бұрын

I agree but Sora is also not out yet, and maybe there are reasons for this. Maybe the computation requirements are still too high for it to be practical. What can competition do if they use way more computation? I think that’s a relevant question to ask. But yes I agree Sora is extremly impressive.

@armondtanz

3 ай бұрын

like when you try on a pair of shiny new shoes in a shop then notice your old shabby shoes lying there...

@LouisGedo

3 ай бұрын

@@armondtanz Yup!

The comparison is very informative! The frame quality of Sora is much better than others. It looks like every frame is MJ-level. Thanks for sharing!

I think your assessment is spot on. Thanks for the content - it's very informative.....

Oddly Leonardo's with facefusion run over the top might actually be useable. Isn't that just SVD1.1? Sora really is just a crazy jump in quality, can't wait to try it out.

@armondtanz

3 ай бұрын

Leonardo caught my eye on the results

Apart from the massive compute scale that seems to set Sora apart, it's seems the be the only model which generates the whole video in one go. That's the key detail that helps it so much with temporal coherence, since the model plans the whole sequence from the beginning, instead of trying to keep track of the last few frames in order to generate the next one...

Thank you very much for this review!!

there are def alot of papers out that talk about other methods like boxamator that give alot more control over the the process - i think we need to see a sort of hybrid boxamator / segmentation/ contronet approach that lets u tag elements in the shot to behave in a certain way - move in a path and control the camera - if we can then prompt per segment to give each group a action we might really get something closer to what Sora is showing in comfy ui

Love the channel! keep it up

Any evaluation of AI has a very short life span. There are still people saying "AI gives you extra fingers" :) Of course next year there will be open source tools better than Sora, and some other company will create the next generation AI video generator, that will crush the competition. At least, we know that by experience.

Not without ControlNets, Dreambooth training, LoRas, and fine controls, I don't believe it will. There's just not enough control with Runway, Pika, SORA or any of these generative models for me. I'm at the point where I'm training my own DreamBooth models/characters, hand-compositing scenes, painting in light, interpolating, ect. I'd love for Runway or SORA to roll these kinds of things in, but I don't see it happening anytime soon, not beyond simple camera moves. And there's also the closed-source/subscription aspect of it that I don't really like. Anyway, keep experimenting man - I've already picked up some things from you and adopted them into my workflow!

Awesome comparison. Thanks! I think the computer power is a factor. But... I'd be willing to wait an hour or more to generate these clips if I can pay less. Actually, I heard many of the 60 second Sora clips took about an hour to render, so maybe I'd have to wait overnight or something.

Pretty sure the model architecture is different a well. The also summarize data across frames into patches . Would be interested to see if Stabe diffusion can catch up at all

@mickmumpitz

3 ай бұрын

I really hope so!

@ChristianIce

3 ай бұрын

Not "if", just "when" ;)

The quality of current models can indeed be pushed further with more compute. As an example, if your machine is beefy enough, and you are patient, you can run Stable Video Diffusion, at e.g. 1920*1080 resolution and 75 steps. It takes around 10-15 minutes to run with GTX4090, but the quality is much better than with default settings.

so Leonrado AI is the best for now to use

I think the large companies with a ton of money to spend on these projects will always have the best quality. Open source models will always lag behind, but continually improve. I'm betting Stable Diffusion 3 will likely be on a similar level as Dall-E 3, but then Sora will become available and then the open source guys will have to play catch up again. As far as video goes, I think there will be a lot of ground to cover to get to the quality Sora is able to produce.

Hello from Argentina! I think that Runway or Pika will bring Sora via an API and we will be using Sora through the environment of Pika or Runway, or InVideo or what you prefer. Today, bloggers use, for example, Jasper to write content which brings OpenAI. They do not use ChatGPT directly. Sorry if the message was too long.

It is truly depressing. Aquiring skills takes years. Making them obsolete takes days for these AIs.

@syedharis3235

3 ай бұрын

Imo not a single skill is obsolete yet.... some of them are minimized to certain extent but not get obsolete

There are rumors that Sora takes a very long time to generate a clip. For example, if it takes 2 hours for a video, then you cannot work properly with AI video.

@mickmumpitz

3 ай бұрын

Then it really would hardly be usable. And if you then consider that it's two hours on absolute super computers... The electricity costs alone must be enormous.

@martin231181

3 ай бұрын

That maybe is why it’s not released yet. Hardware need to catch up.

@phen-themoogle7651

3 ай бұрын

@@mickmumpitz Sam Altman made some videos live while using twitter within a few minutes I think it was so that's not on a super computer most likely. But I don't remember the amount of time between the comments and generations, would be cool to look into if somebody knows or can view them. "given the advancements in AI and its integration into various platforms, it's plausible that these demonstrations could be performed on high-performance workstations rather than requiring supercomputers. OpenAI's models are designed to be accessible and usable, suggesting the use of reasonably powerful but conventional computing resources." Would make most sense if he didn't use something too powerful for just a twitter demonstration if he intends to release it to the public someday since most of us don't have super computers...

They are all going to have to catch up quick to Sora. Hopefully Open AI does not raise their price when this roles out.

*SDXL WORKLFLOW in COMFYUI* ???🥺

I suspect that LeonardoAI is trained using a very similar dataset as Sora.

I have some good results with picverse.

Here is the thing that separates Sora from the others is without editing you can make videos a minute long and so far of what I've seen no other have been able to be longer than maybe 6 seconds

I think that we will experience similar phenomena in video generation as in image generation and text generation: getting more performance out of more compressed models. Simply because AI itself can be applied more and more to explore where the sweet spot of efficiency lies. The more clearly the required principles are "peeled out" for quality - the stronger the "wow" effect will be. But the most exciting thing is whether Yan LeCun will be right that what OpenAI is pursuing with Sora is an aberration because it is not a real "world model". I think we will increasingly realize that there is greater intelligence in the depths of neural networks and much greater and surprising contexts that may yet be unearthed. And that's a bit scary, because this development is continuing at an exponential rate. We are close to the threshold where the tools exceed our biological limitations. And what happens THEN is a question of controllability - in a human global context of power competition. Ui ui ui...

@sudoverse2342

16 күн бұрын

it terrified me using ai with early disco and stable diffusion to animate last year. Because of what you described, the depth of understanding of how reality is built in the neural networks. It is hard to describe, but as an artist, you can feel the power of a god in the way AI can understand the essence of ideas. These new polished interfaces take away that uncanniness of having forbidden knowledge with the AI. It just makes things look "realistic" which is ultimately pointless because if you want something realistic you need only to open your eyes because we live in a "realistic' world that you can see with eyes and touch with the body.

i also suspect its computational power and that openAI is using that to their advantage. someone should run SD on a supercomputer and tell us

in a few more year, the corporate version of this will wipe floor with all those big studio

SORA is miles ahead, think about where SORA will be by Summer? or Winter???

Benchmarks will eventually be "how long does it take to generate photorealistic videos". Yes sora is ahead, but with 8-32x the compute time to generate. With this in mind it looks like Leonardo is the winner actually.

SORA requires massive computer power - I wonder whenn it I will be available for commercial use. If you put 100x more GPUs to Ranway or Pika you may get similar results.

@veenasuresh7054

3 ай бұрын

What are the good things you see? All that glitters is not gold... Each person's essence is etched in their dreams and daily pursuits, especially in their chosen profession. Yet, the looming shadow of AI replacing human roles casts doubt on our future. It's confounding why some fervently champion innovations that jeopardize our fundamental right to earn a living. Beware, the allure of novelty can exact a devastating toll. Remember. Where does your brain kick into gear? Think about it: When photography came along, it took away some painting gigs, but it also made room for loads of new jobs, like making cameras and developing software. Plus, photography itself sparked a whole bunch of fresh career paths. What amazing job chances do you see, super-smart, when you finish AI tasks with just a little prompt? Give me one or two examples. And can you reassure me that these jobs won't be taken over completely by AI in the future? Could you clarify which area sees a big boost in humanity's quality of life?

👋

Nobody knows for sure. My prediction: free platforms turn to subscription, and Sora's hype will break out the minute it becomes adopted in the service of money.

I'm gonna put my money on that sora is gonna be mega expensive and only affordable by big studios

Open source AI enthusiast need if they want to progress things consider creating an AI@home training capability and processing capability akin to SETI@home and folding@home to share GPU cycles with their favorite projects in order to help the projects offset cost and to help them advance forward because we don't want this type of capabilities only in the hands of behemoths and controlled by them. Because then corporate interest and political lobbying will decide what kind of art and topics you may discuss h ck even historical facts are being evaded in the corporate AI systems so I hope the Open Source and open uncensored models keep thriving because the alternative is mildly put quite chilling

SORA = SORA!

Better name for this video would have been "How much better is Sora?"

@mickmumpitz

3 ай бұрын

Good idea, I think I'll try it :)

Sora = modern. All the rest are living in stone age

@user-cy9uz9tk2m

3 ай бұрын

They will catch up soon , it happens all time

WTF sora!!!!!

It's all still uncanny

@sudoverse2342

16 күн бұрын

People are so incredibly uncreative with ai its unreal. people only want to make real things.

Sup

Wow. Thanks for the comparison. After Sora they basically all suck. They will have to do something to compete.

Open Source has Stable Diffusion 3, which can compete with Sora.

Each person's essence is etched in their dreams and daily pursuits, especially in their chosen profession. Yet, the looming shadow of AI replacing human roles casts doubt on our future. It's confounding why some fervently champion innovations that jeopardize our fundamental right to earn a living. Beware, the allure of novelty can exact a devastating toll. Remember. It's baffling why those in power don't halt technologies that threaten lives, and why courts or human rights groups don't step in. Ordinary folks uphold the world's wealth structure, yet their job security is at risk. It's the duty of every government to protect their people's livelihoods.

they're all milesssssssss away from sora lol

sORA IT'S fAKE

Sora get killed.... Watch out for emo its unimaginably realistic and i think there is nothing more left.If we are talking about AI videos,may be there is some kind of more features that user could add 2 face , change dress instantly, could die and rebirth or even time travel 🤣🤣🤣🤣🤣 Things got out of control. Brace ourselves..

Open AI sucks .

This seemed a pointless exercise with results we already would know. Obviously, SORA is superior by many degrees. What is the point in speculating how the others might catch up one day based on speculative causes? It seems a waste of time by the video creator. And also by we viewers.

@blueyc4rter

3 ай бұрын

Still interesting to see the next best alternatives considering that we can't test SORA yet.

Please help.your last video Shows the error Error occurred when executing ADE_LoadAnimateDiffModel: No pos_encoder.pe found in mm_state_dict - sd15_t2v_beta.ckpt is not a valid AnimateDiff motion module! File "D:\Blender_ComfyUI\ComfyUI\execution.py", line 155, in recursive_execute output_data, output_ui = get_output_data(obj, input_data_all) File "D:\Blender_ComfyUI\ComfyUI\execution.py", line 85, in get_output_data return_values = map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True) File "D:\Blender_ComfyUI\ComfyUI\execution.py", line 78, in map_node_over_list results.append(getattr(obj, func)(**slice_dict(input_data_all, i))) File "D:\Blender_ComfyUI\ComfyUI\custom_nodes\ComfyUI-AnimateDiff-Evolved\animatediff odes_gen2.py", line 170, in load_motion_model motion_model = load_motion_module_gen2(model_name=model_name, motion_model_settings=ad_settings) File "D:\Blender_ComfyUI\ComfyUI\custom_nodes\ComfyUI-AnimateDiff-Evolved\animatediff\model_injection.py", line 389, in load_motion_module_gen2 ad_wrapper = AnimateDiffModel(mm_state_dict=mm_state_dict, mm_info=mm_info) File "D:\Blender_ComfyUI\ComfyUI\custom_nodes\ComfyUI-AnimateDiff-Evolved\animatediff\motion_module_ad.py", line 160, in __init__ self.encoding_max_len = get_position_encoding_max_len(mm_state_dict, mm_info.mm_name) File "D:\Blender_ComfyUI\ComfyUI\custom_nodes\ComfyUI-AnimateDiff-Evolved\animatediff\motion_module_ad.py", line 89, in get_position_encoding_max_len raise MotionCompatibilityError(f"No pos_encoder.pe found in mm_state_dict - {mm_name} is not a valid AnimateDiff motion module!")